Speech Errors as a Window on Language and Thought: A Cognitive Science Perspective by Giulia M.L. Bencini 1. INTRODUCTION We are so used to speaking in our native language that, under normal circumstances, we take this ability for granted. We don’t assume that “talking” involves mental operations − “thinking” is where we assume that real cognitive activity resides. Under normal circumstances, native speakers only appreciate the difficulty of executing speech when something goes wrong, such as when they are temporarily unable to find the right word (tip-of-the-tongue state, TOT), or when they utter something different from what they intended to say, i.e. they produce a slip of the tongue. To the lay person, slips of the tongue are likely to be sources of amusement or embarrassment, but to the cognitive scientist they are natural experiments that reveal properties of the system that builds utterances and translates our thoughts into language. This paper reviews the contributions of speech error data to cognitive theories of language production, because errors form part of the empirical data set upon which psycholinguistic theories of production were historically built on. With some notable exceptions (e.g. Fromkin 1971) linguists have not taken speech error data to be Saggi/Ensayos/Essais/Essays Errori / Errors – 04/2017 243 informative about speakers’ linguistic knowledge or mental grammars. The paper wishes to rectify this and place speech errors (and more generally, language production research) back onto the linguistic data map, within a more general commitment to aim for unified theories of language representation and processing (e.g. Bencini 2013; Lewis and Phillips 2015; Branigan and Pickering 2016). What is a speech error? A speech error (also known as slip of the tongue) is operationally defined as a deviation from a speaker’s intended utterance. This definition requires knowing what a speaker’s intended utterance is, which limits the classification of errors to instances in which the intended utterance is uniquely inferable from context, or when it can be verified through other means, such as in experimental paradigms that are specifically designed to elicit errors where the target utterance is either provided or constrained to an identifiable set of alternatives. Errors occur in all languages (including sign languages) and across input and output modality (e.g. reading comprehension, written production), but I will not address these other modalities here. The focus of the paper is on spoken errors made by monolingual English native speakers, the systematic properties of those errors, and an outline of a cognitive architecture (or language production model) that explains how those errors arise during real-time language processing. Following Ferreira and Svlec (2007) I refer to the model as the consensus model for language production. Later in the paper I will review more recent challenges to the consensus model and, in light of additional experimental data, I will present a revised model for language production that aims to unify data from representation and processing. The paper aims to show that: 1) native speaker errors are rich sources of data on the nature of language representation and language use. They are, therefore, primary data for linguistics, 2) speech errors are best understood within language processing models that assemble utterances from their component units and have an overall frame-based (constructionist) architecture where frames are filled with the appropriate linguistic units: words at the syntactic level and phonemes at the phonological level. The paper is organized as follows: In section 2, I provide an overview of the cognitive science perspective on errors. In section 3, I review processing models of language production. In order to understand the properties of speech errors and how they occur, section 4 discusses speaker error types in relation to “when and where” they occur in the processing model. Section 5 reviews some outstanding issues and current debates in language production and section 6 proposes a revised language production model to account for some of these debates. 2. A COGNITIVE SCIENCE PERSPECTIVE ON SPEAKING AND MISSPEAKING The commonsense view of speaking alluded to earlier rests on an implicit assumption that Bock (1996) refers to as the “mind-in-the mouth assumption”. This assumption is also associated with the tradition of behaviorist psychology and linguistics. “Mind-in Saggi/Ensayos/Essais/Essays Errori / Errors – 04/2017 244 the-mouth” refers to an output-oriented view of cognition and the idea that there is a “shallow” relationship between pre-linguistic thought and speech: no intermediate steps are required from, for example, seeing an object to naming it. This view contrasts with the more recent perspective developed in the cognitive sciences for which generating everyday utterances is a prodigious act of linguistic creativity in the face of the computational complexity of the task. We will see, on the basis of error data and data from psycholinguistic experiments, that speaking involves selecting, retrieving and building novel combinations of units at different levels of representation, from semantics to syntax, to morphology and phonology. At each level there are different rules or regularities that speakers unconsciously follow, from rules about word order to rules about word building to rules about the sequencing of sound units. Viewed as information processing (Bock 1982), speaking is a feat of computational efficiency. The mental lexicon for the average educated adult contains anywhere between 50,000 and 100,000 words and utterance generation happens fast: during normal fluent speech, speakers can produce up to three words per second (Levelt 1999). Investigators have collected speech error corpora through direct observation, either by listening to speech and systematically writing down all the errors detected (e.g. Meringer and Meyer 1895), or by setting aside “collection times” during the day, or, more systematically, by recording speech for later examination (Garnham et al. 1981). Using this technique, Garnham et al. (1981) were able to estimate that errors in normal native adult English speakers occur at a rate of one or two errors every thousand words. 2.1 Language production data as data for linguistics A complete cognitive theory of human language requires an account of the representations and the processes that explain how it is that humans represent and use (comprehend and produce) language. In practice, linguists in the generative tradition have proposed candidate representations (in particular with respect to syntactic and semantic structure) largely relying on explicit judgments about the grammatical (or semantic) acceptability of individual sentences – henceforth, acceptability judgments. Over the last 25-30 years, alternative models of linguistic knowledge have developed within a class of theories that can be collectively referred to as usage-based linguistic approaches (following Langacker 1988), including Cognitive Linguistic Approaches and the family of Construction Grammars (for recent overviews, see the edited volumes by Dabrowska and Divjak (2015) for the former; Hoffmann and Trousdale (2013) for the latter). Although usage-based approaches differ in many important details, they share two important philosophical and methodological characteristics: A “cognitive commitment” to converge with data from experimental psychology, cognitive science and neuroscience and a broadened empirical basis, where (at least in principle) psycholinguistic data are to be placed on equal footing with traditional linguistic data. The revised model in section 6 is an example of a model Saggi/Ensayos/Essais/Essays Errori / Errors – 04/2017 245 that integrates theoretical and experimental developments within a cognitive constructionist approach to argument structure (Goldberg 1995; 2006) and a modified psycholinguistic language production model. The model is also compatible with otherwise very different linguistic neo-constructionist approaches (e.g. Borer 2005; Levin and Rappaport Hovav 2005, i.a.) Psycholinguists over the last 50 years have examined the properties and distributions of naturalistically occurring errors (in error corpora) and set up experimental procedures to make errors more likely (e.g. Baars, Motley and MacKay 1975; Baars 1992; Warldlow Lane and Ferreira 2010). In one such technique, called SLIP (Spoonerisms of Laboratory-Induced Predisposition), speakers are presented with pairs of words on a computer screen and are asked to read them silently. Every few trials, on a tone, speakers are cued to produce the most recent word pair as quickly as possible, e.g. “barn door”. By skillfully setting up the preceding word pair sequence, (e.g. “dog bone”, “dust ball” “dead bug” “deer back” “doll bed”) the phrase “barn door” on occasion will be produced erroneously as “darn bore”. Nature breaks at its joints: the types and distributions of errors reveal how the system is built to begin with. In the error where the speaker uttered “darn bore” instead of “barn door”, there was an exchange between two word initial sounds: /b/ and /d/. The experimental set-up induced higher rates of speech errors than can be typically observed in natural speech corpora, but importantly, the mechanism is “normal” speech production (Bock and Levelt 1994). The error reveals a number of interesting properties about the system that builds words and utterances, among which that